To address the challenges of unbalanced class labels with varying maturity levels of tomato fruits and low recognition accuracy for both fruits and stems in intelligent harvesting, we propose the YOLOX-SE-GIoU model for identifying tomato fruit maturity and stems. The SE focus module was incorporated into YOLOX to improve the identification accuracy, addressing the imbalance in the number of tomato fruits and stems. Additionally, we optimized the loss function to GIoU loss to minimize discrepancies across different scales of fruits and stems. The mean average precision (mAP) of the improved YOLOX-SE-GIoU model reaches 92.17%. Compared to YOLOv4, YOLOv5, YOLOv7, and YOLOX models, the improved model shows an improvement of 1.17–22.21%. The average precision (AP) for unbalanced semi-ripe tomatoes increased by 1.68–26.66%, while the AP for stems increased by 3.78–45.03%. Experimental results demonstrate that the YOLOX-SE-GIoU model exhibits superior overall recognition performance for unbalanced and scale-variant samples compared to the original model and other models in the same series. It effectively reduces false and missed detections during tomato harvesting, improving the identification accuracy of tomato fruits and stems. The findings of this work provide a technical foundation for developing advanced fruit harvesting techniques.

Apple leaf diseases significantly impair the photosynthetic efficiency and growth quality of apple trees, leading to reduced fruit yields. Existing methods for disease detection and severity classification struggle to quickly and accurately segment and quantify diseased areas on leaves, particularly in complex backgrounds. To address this issue, we propose a method for assessing the severity of apple leaf diseases based on a combination of improved HRNet and DRL-watershed algorithms. First, we selected HRNet_w32 as the backbone feature extraction network and incorporated a Normalization Attention Mechanism (NAM). Then, we combined the Dice Loss and Focal Loss functions to construct an enhanced HRNet based semantic segmentation model for pixel-level segmentation of both apple leaf and diseased regions. Furthermore, the segmented leaf and disease regions were further optimized using the DRL-watershed algorithm to distinguish overlapping leaf regions. Experimental results demonstrate that the modified HRNet model achieved a mean intersection over union (mIoU) of 88.91% and a mean pixel accuracy (mPA) of 94.13%, representing improvements of 8.77 and 7.25% points, respectively, over the original HRNet. The disease severity assessment accuracy reached 97.65%. This study not only accurately segments apple leaves and diseased areas, but also effectively addresses the impact of complex backgrounds and leaf overlap on disease severity assessment, providing a solid scientific basis for disease management strategies.

Due to the complex environment of pig farms and the diversity of pig behaviors, the existing methods based on deep learning cannot meet the requirements of high accuracy and real-time detection of pig eating behavior. In this paper, a SE-YOLOX model for detecting pig diet and drinking behaviors was designed. In this model, a Squeeze-and-Excitation (SE) attention module is introduced between the neck layer and the prediction layer of YOLOX, and the input feature map is compressed into a vector through global average pooling operation, and then mapped to a smaller vector through a fully connected layer. A sigmoid function is also used to compress each element in this vector to between 0 and 1 and multiply it with the original input feature map to get the weighted feature map. Through SE attention mechanism, the model can learn the importance of each channel adaptively, thus improving the detection accuracy. The experimental results show that the mean Average Prediction (mAP) of the SE-YOLOX model is 88.03%, which is higher than 13.11% of the original YOLOX model. SE-YOLOX can ensure real-time performance, it also can effectively improve the accuracy of pig diet and drinking water behavior detection.

Recommender systems are vulnerable to poisoning attacks due to their open nature, allowing attackers to inject malicious user profiles to deliberately manipulate recommendation results. Existing detection methods mainly focus on rating behaviors while neglecting key semantic information such as item labels, making them ineffective in handling complex or highly camouflaged attacks. To address the limitations of existing detection methods, which overly rely on ratings and insufficiently exploit the semantic association information contained in item labels, we employ a pretrained large language model to encode label semantics and fuse rating information with label semantic information to jointly identify malicious users, thereby proposing PAD-LLM, a poisoning attack detection method based on large language model-based label semantic encoding. First, we adopt a Text-to-Text Transfer Transformer model to semantically encode label text sequences and fuse them with rating behaviors to construct a user-item-label three-dimensional tensor representation, thereby enabling unified modeling of multi-source heterogeneous data. On this basis, we design a local–global joint feature extraction framework, in which three-dimensional depthwise separable convolution and multi-head Performer are used to jointly model local interaction patterns and global dependency structures, while a gated residual mechanism is introduced to achieve dynamic fusion. Furthermore, we incorporate contrastive learning to enhance the inter-class separability of latent representations, thereby improving the identification capability for malicious user profiles. We conduct comparative experiments on the MovieLens-1 M and Amazon datasets. The results demonstrate that PAD-LLM achieves better detection performance than baseline methods under multiple poisoning attack settings.

Recommender systems are vulnerable to poisoning attacks due to their open nature, and attackers can inject malicious user profiles to deliberately manipulate the recommendation results. Existing detection methods mainly focus on rating behaviors while neglecting key semantic information such as item labels, making them ineffective in handling complex or highly camouflaged attacks. To overcome the limitations that detection methods overly rely on ratings and insufficiently exploit semantic association information in item labels, we use a pretrained large language model to encode label semantic information, and fuse rating information with label semantic information to jointly identify malicious users, thereby proposing a poisoning attack detection method based on large language model–based label semantic encoding, PAD-LLM. First, we adopt a text-to-text Transfer Transformer model to semantically encode the label text sequences, and fuse them with rating behaviors to construct a user-item-label three-dimensional tensor representation, thereby enabling unified modeling of multi-source heterogeneous data. On this basis, we design a local-lobal joint feature extraction framework, via three-dimensional depthwise separable convolution and multi-head Performer to jointly model local interaction patterns and global dependency structures, and via a gated residual mechanism to realize dynamic fusion; furthermore, we incorporate contrastive learning to enhance the inter-class separability of latent representations, thereby improving the identification capability for malicious user profiles. We conduct comparative experiments on the MovieLens-1M and Amazon datasets. The results demonstrate that PAD-LLM achieves better detection performance than the baseline methods under multiple poisoning attack settings.